Category-specific general Pavlovian-instrumental transfer

Modern living is characterized by easy access to highly palatable energy-dense foods. Environmental cues associated with palatable foods increase seeking of those foods (specific transfer) and other palatable foods (general transfer). We conducted a series of studies testing the boundaries of food cue-reactivity by evaluating the impact of broader flavor associations (i.e. saltiness, sweetness) in eliciting general transfer effects. Experiment 1 was an online experiment with fictive rewards that tested if two actions associated with different food rewards (chip and chocolate points) could be provoked by images of other foods that were either similar or distinct in flavor from the foods associated with these instrumental actions. We observed that response excitation was only elicited by similarly flavored food cues, whereas distinctly flavored food cues inhibited response rates relative to control cues. Experiment 2 confirmed this observation in a classroom setting where real food rewards were contingent on task performance. Experiment 3 was an online study that further confirmed the reliability of the effects with a well powered sample. There were moderate-to-strong associations between specific and general transfer effects across all studies, suggesting overlapping cognitive processes are responsible for both transfer effects. These data improve the mechanistic understanding of how broad category associations can moderate the impact of food cues on food choices. This knowledge could be helpful for improving the precision of psychological interventions that seek to mitigate the impact of food cue-reactivity.


Introduction
Modern living often involves access to highly palatable energy-dense foods (Berthoud, 2012).Environmental cues become associated with several features of food rewards (i.e. the specific flavor, smell, and texture profile), and come to trigger the wanting of highly palatable foods (Berridge, Ho, Richard, & DiFeliceantonio, 2010;Boswell & Kober, 2016).These features, as well as other abstract features, such as the membership of a broader category of rewards (e.g."salty snacks") may bias the likelihood of certain actions (Hommel, 2022;Hommel, Müsseler, Aschersleben, & Prinz, 2001), such as choosing highly palatable foods over healthier alternatives (Kirsten, Seib-Pfeifer, & Gibbons, 2022;Watson, Wiers, Hommel, Ridderinkhof, & De Wit, 2016).Given that food cue-reactivity is predictive of food consumption and weight gain (Boswell & Kober, 2016), it is important to clarify the mechanisms that underpin food cue-reactivity.In so doing interventions for obesity and eating disorders can be appropriately targeted to combat the impact of the obesogenic environment.
The psychological processes underpinning cue-induced changes in food choice have be studied using the Pavlovian-instrumental-transfer (PIT) paradigm (for a review, see Cartoni, Balleine, & Baldassarre, 2016).The PIT paradigm has contributed to understanding the mechanisms of cue-reactivity in animals and humans, and specifically the role of cue-reactivity in obesity (Kanoski & Boutelle, 2022).In the experimental paradigm, instrumental actions (e.g.actions for seeking chip and chocolate outcomes) are typically trained separately to the Pavlovian cues (e.g.cues that predict chip or chocolate outcomes in the absence of instrumental actions).The impact of Pavlovian cues on instrumental actions is evaluated in a transfer test, where for the first time, trained actions are available during intermittent exposure to the Pavlovian cues.There are typically two kinds of transfer effects: specific and general transfer.Specific transfer involves a reward cue (e.g. a cue associated with potato chips) increasing rates of responding for a congruent reward-seeking action (e.g. an action associated with potato chips) more than an incongruent reward-seeking action (e.g. an action associated with chocolate).Evidence for specific transfer consistently shows that reward cues robustly invigorate congruent reward-seeking actions for a variety of different reward types (e.g.Alarcón & Delamater, 2019;Leung & Balleine, 2015;Mahlberg, Weidemann, Hogarth, & Moustafa, 2019;Quail, Morris, & Balleine, 2017;Watson, Wiers, Hommel, & de Wit, 2014).
A sensory-specific account of specific transfer suggests that the reward cue motivates responding for a congruent reward via their association with the perceptual properties of the reward (i.e. the identity of the reward), and its desirability (i.e., incentive salience) is unimportant (Cartoni et al., 2016;Watson & de Wit, 2018).Some evidence converges with the sensory-specific account, showing that manipulations of food desirability (e.g.making the food available until the subject reaches satiety) do not produce significant changes in specific PIT, consistent with the sensory-specific account (e.g.Corbit, Janak, & Balleine, 2007;Hogarth & Chase, 2011;Holland, 2004;Watson et al., 2014).However, other evidence is inconsistent with the sensory-specific account, as it shows that specific transfer effects are impacted by shifts in outcome desirability in some circumstances (Allman, DeLeon, Cataldo, Holland, & Johnson, 2010;Hinojosa-Aguayo & González, 2020;Seabrooke, Hogarth, Edmunds, & Mitchell, 2019), even though specific transfer effects persist when food desirability is reduced (Lingawi, Berman, Bounds, & Laurent, 2022;Sommer, Münster, Fehrentz, & Hauber, 2022).This evidence is consistent with a goal-directed account of specific transfer, whereby the Pavlovian cue primarily motivates congruent reward-seeking actions via their mutual association with (1) the perceptual identity of the reward, and (2) the incentive salience of the (i.e. its desirability; for a review of the evidence see Mahlberg, Seabrooke, et al., 2019).
General transfer involves Pavlovian cues increasing instrumental actions for a reward different from the cue it was associated with (e.g. a cue associated with cashew nuts while responding for popcorn and chocolate, see Watson et al., 2014).The available evidence shows that in some circumstances reward cues increase the response rate for other reward-seeking actions (Belanger et al., 2022;Corbit & Balleine, 2005, 2011;Corbit et al., 2007;Corbit, Fischbach, & Janak, 2016;Laurent, Priya, Crimmins, & Balleine, 2021;Lingawi et al., 2022;Pielock, Lex, & Hauber, 2011;Quail et al., 2017;Watson et al., 2014).Theoretical accounts of general PIT differ in the extent to which they suggest this depends on the perceptual properties of a reward.An energizing account of general transfer suggests that the reward cues motivate responding for all reward-seeking actions via the cue's association with the incentive salience of the reward (i.e. its desirability).Research in non-human animals has shown that shifts in desirability of the reward associated with the cue modulate the strength of the general transfer (Corbit et al., 2007).In human translation studies, the evidence for general transfer and the energizing account is less compelling.Some human studies have observed positive evidence of general transfer in the presence of reward cues (Quail et al., 2017;Watson et al., 2014).However, several other attempts to observe general transfer in humans have been unsuccessful (Belanger et al., 2022;Geurts, Huys, den Ouden, & Cools, 2013;Meemken & Horstmann, 2019;Nadler, Delgado, & Delamater, 2011).
Inconsistent evidence for general transfer in human participants, combined with evidence showing that specific transfer is sensitive to reward value, suggests that both the perceptual identity and the incentive salience of the reward are important influences in the cue biasing action.A goal-directed account of general PIT suggests that reward cues motivate reward-seeking actions via incentive salience (i.e., desirability), and associations with perceptual properties of reward (i.e., the identity) play a role in producing general PIT effects.This account predicts that reward cues will motivate reward-seeking actions for other rewards, but (unlike the energizing account) also predicts that this effect will be selective, producing greater responding for rewards which share perceptual similarities with the reward the cue was associated with than for rewards which are more dissimilar.For example, a fast-food advertisement (e.g., for a Pizza restaurant) may prompt actions to seek rewards with overlapping sensory qualities (e.g., retrieving a salty snack like potato chips) compared to other rewards (e.g., sweet snacks like chocolate).There is precedence for selective cue-reactivity effects: studies have shown that food cues will increase desire for foods from a similar category, but not from a dissimilar category.For example, Ferriday and Brunstrom (2011) found that exposure to the sight and smell of highly palatable savory food (pizza) cued craving for other savory foods (pasta, eggs, chips and beans, and chicken tikka masala), but not for palatable sweet foods (cake and chocolate buttons).
With this in mind, a reward cue (e.g. a cue for ice cream) used in a general PIT transfer test might bias food choice via its broader associations (e.g. its association with 'sweet snacks'), particularly if one of the available choices has congruent features (e.g. a response associated with chocolate; (Hommel, 2022;Hommel et al., 2001).The influence of shared associative connections may be more pronounced in general PIT tasks that utilize a choice task due to competition between reward choices (e.g., potato chips and chocolate).Competition might encourage utilizing shared associations between the cue and the action to select the optimal action (Alarcón & Delamater, 2019;Corbit & Janak, 2007).However, it is currently untested whether category-based action selection is a reliable moderator of general transfer effects, as studies thus far have not been designed to test this.It is of critical importance that this is tested, as evidence of category-based action selection (and the absence of non-specific response elevation) during general transfer would support a goal-directed view and contradict the prevailing energizing hypothesis of general PIT.
Therefore, in the current study, we conducted three experiments to test the energizing and goal-directed general PIT hypotheses, specifically examining the extent to which category-based shared associations between the reward cue and available outcomes influence rates of responding for those outcomes.We preregistered our hypotheses with the open science framework (https://doi.org/10.17605/OSF.IO/ U4G7P).Experiment 1 tested general transfer on trained rewardseeking responses (for potato chips and chocolate points) using two distinct reward cues (popcorn and ice-cream images) compared to a control cue not associated with reward (a blank white image).The experiment was conducted in an online student sample.An energizing account of general PIT effect would predict both reward cues would increase response rates from both available actions to a similar extent, compared to the control cue.A goal-directed account of general PIT would predict that reward cues will selectively increase responding for rewards which share category membership with the reward associated with the cue (e.g., the popcorn image eliciting greater response rates for the potato chip action).Specific transfer was also tested in all experiments to confirm the paradigm produced specific transfer, and to explore whether there was any relationship between the extent of specific and general transfer.Specific transfer was assessed by comparing the response rate for a reward (e.g., a chocolate response) during exposure to a cue associated with the same reward (e.g., a chocolate image) and a cue associated with a different reward (e.g. a potato chips image).Responding was tested in a choice condition (i.e., both rewarded responses were available together), and in an "alone" condition where the only other available response was one that never produced a reward.Experiment 2 tested an in-class student sample to confirm the reliability of the results of Experiment 1 in a paradigm where the reward points earned during the PIT task resulted in obtaining food rewards.Experiment 3 aimed to confirm the findings using a similar design to Experiment 2 in a well-powered sample with small design tweaks to strengthen instrumental contingency knowledge.

Participants
Participants were undergraduate students from Western Sydney University enrolled in psychology subjects.There were no exclusion criteria other than the requirement that the task be completed on a computer with a mouse and keyboard.These studies were approved by the Western Sydney University Human Research Ethics Committee (approval number: H14572), and all participants included in the analyses described provided written consent for their data to be used for research purposes.Students participated in these experiments as part of their course work.As such, we allowed students to participate, so they could attain the learning benefits associated with research participation, but we allowed them to opt-out from having their data used for research in order to ensure that consent was voluntary.

Experiment 1
Prior general transfer studies had detected a medium or larger effect size (e.g.Alarcón & Bonardi, 2020;Nadler et al., 2011;Quail et al., 2017), therefore our goal was to reach a sample size powered to detect at least a medium effect.One hundred and seven first-year psychology students volunteered to participate in return for course credit.A total of 70 participants were removed before analysis: nine participants did not consent to their data being used for research purposes; eleven participants showed evidence of distraction during the test phase (the browser tab with the experiment was pushed to the background of their computer, indicating a participant was not focused on the stimuli or able to respond, and therefore the ability for the data to meaningfully contribute to whether stimuli changed their responding was questionable); one subject did not complete the entire experiment; 49 subjects (56.98%) did not correctly answer the R-O knowledge questions after the test phase and were excluded from further analyses; leaving a total of 37 participants in the final sample.Power simulations suggested that the final sample had sufficient power for detecting the effects of interest (see the supplementary document).

Experiment 2
Seventy-six second-year psychology students participated in Experiment 2 as an in-class activity at Western Sydney University, therefore our sample size was dependent on the number of students who attended class.A total of 20 participants were removed before analysis: 12 participants indicated they did not consent to have their data used for research purposes, so these data were removed.Two people were removed as they showed distraction during sections of the test phase.Six people completed the study online, which meant that they did not have access to the food prizes, so they were excluded from the analysis.A further 22 (28.95%) did not correctly answer the R-O knowledge questions after the test phase and therefore were suppressed from analyses.This left a total of 34 participants in the final sample.Power simulations showed that the sample had sufficient power for detecting the effects of interest (see the supplementary document).

Experiment 3
A power simulation based on the average cue-elicited responding during the general transfer test of Experiment 2 indicated that 95% power could be achieved for the pairwise contrasts with a sample of ~94 participants.Therefore, 133 psychology students at Western Sydney University were recruited to participate in Experiment 3 for course credit.A total of 42 participants were removed before analysis: 17 participants did not consent to their data being used for research purposes; 3 participants showed evidence of distraction during the test; 22 subjects (19.47%) did not correctly answer the R-O knowledge questions after the test phase and were excluded from further analyses; leaving a total of 91 participants in the final sample, which provided sufficient power to confirm prior observations from Experiment 1 and 2. Power simulations suggested that the final sample had sufficient power for detecting the effects of interest (see the supplementary document).

Pavlovian-instrumental transfer (PIT) paradigm
All experiments used a PIT paradigm which included an instrumental training phase, a transfer test phase, and response-outcome tests after each phase, a standard design for testing Pavlovian transfer effects (Mahlberg, Seabrooke, et al., 2019).See Fig. 1 for a visual overview and Table 1 for a further procedural summary, and we describe each phase of the task below.The task was programmed using jsPsych (de Leeuw, 2015) and hosted online using Pavlovia.org.The experimental code can be accessed through the open science repository for this manuscript.The experiments required a computer with a mouse.For Experiment 1, the experimental procedure was accessed remotely due to COVID-19 lockdowns.For Experiment 2, the task was accessed during a tutorial, and food rewards (potato chips and chocolate) were provided to students at the end of the experiment according to the points earned.The food rewards were Cadbury Dairy Milk Share packs (chocolate) and Smith's Crinkle Cut multipacks (chips).As with Experiment 1, Experiment 3 was accessed remotely.

Instrumental training.
Participants could win chip and chocolate points by clicking green buttons on the left or right of the screen using their mouse.Each response was exclusively paired with one reward outcome.For example, the left button (i.e., R salty ) always retrieved potato chip points and the right button (i.e., R sweet ) always retrieved chocolate points.These response-outcome pairs were counterbalanced, thus for some participants, the left button was paired with chip points and the right button with chocolate points, while other participants experienced the opposite associations.R salty and R sweet were each partially reinforced on a random ratio schedule (as in Quail et al., 2017) so that each response required an average of five responses to retrieve their paired outcome.Points earned were displayed at the top of the screen above the buttons.When a point was awarded, there was a 2 s time-out during which the reward buttons were not available, as denoted by grey appearance and the lack of animation when it was clicked.Experiment 1 included a middle button (R 0 ) which was not paired with a reward, but ensured all blocks included two responses and served as a comparison response for the rewarded responses.Experiment 1 included three blocks of instrumental training, where participants were required to earn 10 reward points in total to proceed to the next block.One of the training blocks was a 'choice' block, with the two reward-associated responses available, and the middle button was unavailable.There were also two 'alone' blocks, where one of the rewarded responses was available with R 0 , while the other was unavailable (grey and non-clickable).The purpose of the alone blocks was to capture the transfer effects in a condition where only one reinforced response was available to compare to responding in the "choice" blocks where both reinforced responses were available in order to assess whether competition between responses moderates the transfer effects (as argued by Alarcón & Delamater, 2019;Corbit & Janak, 2007).For Experiment 2, there was no middle button, and the choice block was experienced twice, such that a total of 20 reward points had to be earned to exit training and proceed to test.To ensure stronger R-O learning, Experiment 3 followed the design of Experiment 2, but doubled the required points that needed to be earned in each training block (20 points each block, 40 points in total) to exit training.Experiment 3 also differentiated the color of the two available buttons (one was blue and the other was green) to improve discrimination during learning and thus boost R-O knowledge.See Table 1 for a summary of the instrumental training phase for the experiments.

Transfer test.
The transfer test(s) were like the training blocks, with participants able to engage in responses freely.However, the transfer test(s) were under nominal extinction, so the outcome of responses was not known, and their total point counts were hidden and revealed to the participant at the end of the experiment.This ensured that no further instrumental learning took place during the transfer test (s).Responses were recorded throughout, and points were calculated based on the partial reinforcement schedule participants experienced during training (i.e.five responses generated one point for the paired outcome).Stimuli were presented randomly throughout the transfer test.Stimuli were presented for 6 s, with a 4-s interval between each stimulus that captured baseline responding prior to stimulus exposure.Two stimuli had congruent instrumental responses (S congruent ) which were simultaneously incongruent for the other instrumental response (S incongruent ): potato chip stimuli, which were congruent with the potato chip response and incongruent with the chocolate response, and chocolate stimuli which were congruent with the chocolate response and incongruent with the potato response.There was a control cue that was a white square (S-).Two stimuli were from a similar (S similar ) category in their sensory profile to the outcome of one of the instrumental responses, which were simultaneously from a distinct category (S distinct ) category for the other instrumental response: popcorn stimuli (more similar in its flavor category to potato chips, and more distinct in flavor category from chocolate) and ice-cream stimuli (more similar in its flavor category to chocolate, and more distinct in flavor from potato chips).In Experiment 1, there were three transfer test blocks: a 'choice' block, where both rewarded responses were available, and two 'alone' blocks where each of the rewarded response was available as the only rewarded response, and the other response available was R 0 which was never associated with winning points.In Experiment 2 and 3 there was only one transfer test block (equivalent to the 'choice' block in Experiment 1).For all experiments, each block included 15 trials in total, with 3 of each image type randomly presented (chips, chocolate, popcorn, ice-cream, and blank images).For Experiment 1, these 15 trials were repeated for each block (trial order randomized for each block).

R-O knowledge test.
At the end of the instrumental training as well as after transfer test(s), participants were presented with a series of questions to test knowledge of the response-outcome relationships.There were questions about all responses (i.e., three for Experiment 1, two for Experiments 2 and 3).Participants were asked to click the button that earned X outcome (chips, chocolate, or nothing for Experiment 1; chips or chocolate for Experiments 2 and 3).The question order was

General questionnaire
For all studies, the general questionnaire included questions regarding age (in years) and gender (text response).Experiments 2 and 3 also included questions about the location of the experiment (classroom or online), and baseline levels of hunger and thirst using a visual analogue scale ranging from 0 (not at all hungry/thirsty) to 100 (very hungry/thirsty).

Food images
The food images used as reward cues were color images depicting salty snacks (potato chips, popcorn) and sweet snacks (chocolate, icecream).Three unique images were used for each reward type in all experiments (i.e. three each of potato chips, chocolate, popcorn, icecream) and a white square was shown for the three S-trials in each block.The images were sourced from https://unsplash.com.

Procedure
The experiment began by providing participants with an information sheet outlining what they will be expected to do in the experiment, and then presented a consent form.The consent form included an optional checkbox to indicate they wanted to complete the experiment, but that they did not want their data used for research purposes.This was followed by instruction screens explaining how the experiment worked where participants were told that they could earn chips and chocolate points.For Experiments 1 and 3, participants were told that they should imagine that they were in a classroom, and the students who top the class in points for each reward will win those snacks to eat.For Experiment 2, participants were instructed that earning 40 points would earn one of the associated snacks (e.g., earning 40 chip points would earn a snack-sized packet of potato chips).These rewards were available in the classroom for all students that reached this threshold.Following the instruction screens, participants were questioned to ensure they understood the instructions.Participants were required to answer all questions correctly before they could continue; if they answered any of the questions incorrectly, participants were provided with a feedback screen advising them of this and were looped back to the instruction screens.All experiments involved an instrumental training phase followed by a transfer test.Finally, participants were asked to fill out the general questionnaire, and once they did so they were presented with a screen that detailed the total reward point count for each reward (chips and chocolate).

Statistical analysis
R (version 4.2.2.) was used to prepare, visualize, and analyze all data.Below, we describe the data preparation and analyses for manipulation checks we completed to assess for response preferences across the training and transfer test(s) for all experiments ('Response allocation') and manipulation checks to confirm that the samples in each experiment demonstrated learning during training ('Response-outcome knowledge').We also conducted analyses for specific and general transfer test ('Transfer tests') based on pre-registered hypothesis tests, and additional exploratory correlations to assess overlapping variation between specific and general transfer effects we observed ('exploratory analysis').

Response allocation
As a manipulation check, we examined how responses were allocated during training to examine if there were response preferences biasing choice (i.e. between R sweet and Rsalty ); for all experiments we calculated total responses during the training phase and computed paired t-tests.
Similarly, for all experiments we also assessed if responses (R sweet and R salty ) were allocated similarly overall (irrespective of stimulus type) during the transfer tests.In Experiment 1, we assessed with analysis of variance and holm-corrected pairwise comparisons if either rewarded responses (R sweet and R salty ) were preferred compared to each other (i.e.evidence of response bias), as well as if they were preferred over the nonrewarded response R 0 to check the assumption that participants preferred reinforced responses compared to the non-reinforced response.For Experiments 2 and 3, we used a paired t-test to assess for response bias between R sweet and R salty .Additionally, for all experiments, we e additionally calculated paired t-tests to specifically check for biased responding for the reward responses (R sweet and R salty ) were with respect to responding during S-.

Response-outcome knowledge
Response-outcome questions were scored as correct/incorrect for all responses.For all experiments, data from participants who had incorrect response-outcome knowledge for the rewarded responses (R sweet and R salty ) after the transfer phase was not included in the analysis (as in Mahlberg, Weidemann, et al., 2019;Seabrooke et al., 2019;Watson et al., 2014) but see the supplementary for analyses that show including these participants do not change the pattern of results.To show that the included participants learned the response-outcome associations during training, response-outcome knowledge after training was tested using a binomial generalized linear model with a logit link to assess the likelihood of correctly answering response-outcome questions for each outcome (chips, chocolate, nothing).All experiments ascertained whether response-outcome knowledge differed as a function of response type by testing the likelihood of correct answers to the response-outcome knowledge questions as a function of response type using a generalized linear model with a logit link, assessed with analysis of variance using the car package in R. Experiment 1 included three types of responses in the analysis (R sweet , R salty , and R 0 ) and Holm-corrected pairwise-comparisons were calculated to follow up on main effects (Holm, 1979), while Experiments 2 and 3 included two types (R sweet and R salty ).As Experiment 3 also had two blocks of training with response-outcome knowledge tests after each block, we assessed how knowledge changed across the two training blocks in addition to the effect of response type.

Transfer testsdata preparation
Response data was segmented into response rates during stimulus exposure (response rate during S congruent , S incongruent , S similar , S distinct , and S-) and during the intertrial interval preceding stimulus onset (which served as a baseline of response rate for each condition).As baseline response rate can impact the ability to observe cue-elicited responding (Colagiuri & Lovibond, 2015), all models assessing transfer effects included baseline responding as a covariate.For Experiment 1, data was additionally segmented by transfer test block: responses from the blocks where there was a single rewarded response were collapsed to serve as a single comparator to the choice test block.As the R 0 condition was only present for the "alone" condition of the test block contrast it was excluded from these models to avoid issues with model fit.

Transfer testsmodel structures
We attempted to fit all models with mixed effects structures, including random and fixed effects (Matuschek, Kliegl, Vasishth, Baayen, & Bates, 2017).For all Experiments, the response rate data was severely skewed and zero-inflated, which indicated that typical mixed linear modelling was inappropriate.As such, we fit generalized linear mixed models using the glmmTMB package with an additional zero-inflation intercept model to control for zero inflation (Brooks et al., 2017).All models (except for the specific transfer in Experiment 2) included a random intercept for subject to adjust for general variation in responding between participants, and a random slope for response type (R sweet vs. R salty ) to adjust for variation in a participant's response preference.The specific transfer model for Experiment 2 was the exception, as a mixed linear model showed convergence problems and, when simplified, showed near zero random variance across subjects, indicating modelling random effects was inappropriate.Thus, we only modelled fixed effects for this analysis using a standard linear regression approach.
For Experiment 1, specific transfer was assessed with a 2 (stimulus, S congruent vs. S incongruent ) by 2 (test block, choice vs. alone) fixed effects model.Likewise, general transfer tests assessed the response rate in a 3 (stimulus, S similar compared to S-and S distinct ) by 2 (test block, choice vs. alone) fixed effects model.For Experiment 2 and 3, specific and general transfer were both simply tested for their respective fixed effect of stimulus.We report the full model details in the supplementary document, including a comparison between the models reported below and standard linear models to show their superior fit to the data, and to also show that the pattern of results hold regardless of model choice.

Transfer testshypothesis testing
All models were subjected to Analysis of Deviance using the Anova function using Type III sum of squares from the car package for chisquare tests (χ 2 ) for omnibus main effects and interactions, which was the only available method for significance testing fixed effects derived from glmmTMB models (Brooks et al., 2017).Planned contrasts derived from these models tested our pre-registered hypotheses using the emmeans package to generate paired t-tests and p values calculated with Satterthwaite approximation (Luke, 2017;Schad, Vasishth, Hohenstein, & Kliegl, 2020), Holm-corrected for multiple testing (Holm, 1979), and standardized effect sizes (Cohen's d).Given our exclusion criteria resulted in removing a large portion of participants in some experiments, we provide a supplementary document where we present the same analyses we report below, but with all participants included, to show that the pattern remains consistent regardless of whether these participants are excluded.We also report supplementary analyses that show, for experiments 2 and 3, general transfer effects reported below persist when controlling for individual differences in hunger and thirst.

Exploratory analysis
For all experiments, we explored how strongly specific transfer effects (i.e., the difference between responding during congruent and incongruent cues) correlated with the category-specific general transfer effects (i.e., the difference between responding during S similar and S di- stinct ) to gain a preliminary understanding of whether there was some overlap in the underlying processes.

Participants
See Table 2 for average age, gender, and task completion time for both experiments and hunger and thirst data for Experiments 2 and 3.

Experiment 1
Participants made similar numbers of the chip and chocolate responses across the training phase, t(36) = .74,p = .47,d = .18,and points earned for chips did not differ significantly from those earned for chocolate, t(36) = .53,p = .60,d = .17.Similarly, throughout the test phases, there was no evidence for differences in the rate of responses across response type, t(36) = .32,p = .75,d = .06.
During the test phase, participants also demonstrated a general preference for the reinforced responses (R sweet and R salty ) compared to the non-reinforced response (R 0 ), but no preference between R sweet and R salty .There was a main effect of response type, F(2, 108) = 15.07,p < .001.Response rates were similar between the reinforced responses (R sweet and R salty ), t(108) = .27,p = .78,d = .06.However, R sweet response rates were higher than R 0 , t(108) = 4.89, p < .001,d = 1.01, and R salty response rates higher than R 0 , t(108) = 4.61, p < .001,d = .96.There was no significant response bias during the control cue, as R salty showed similar response rates to R sweet during S-, F(1, 72) = .001,p = .97.

Experiment 2
Participants made similar numbers of chip and chocolate responses across the training phase, t(33) = 1.73, p = .09,d = .55,however, significantly more chocolate points, compared to chip points, were earned, t(33) = 2.22, p = .03,d = .76.During the test phase, there was no evidence for differences in the rate of responses across response types overall, t(33) = 1.04, p = .30,d = .33.There was no significant response bias during the control cue, as R salty showed similar response rates to R sweet during S-, F(1, 66) = 3.85, p = .054.

Experiment 3
Participants made similar numbers of chip and chocolate response across the training phase, t(90) = 1.89, p = .063,d = .35,and a similar number of chocolate points were earned compared to chip points, t(33) = 1.31, p = .19,d = .27.However, there was evidence for higher responding for chocolate compared to chips during the test phase, t(90) = 2.17, p = .03,d = .33.There was no significant response bias during the control cue, as R salty showed similar response rates to R sweet during S-, F(1, 180) = .89,p = .35.

Experiment 1
Participants demonstrated similar understanding of the responseoutcome contingencies as a function of response type (R sweet , R salty, and R 0 ).While there was a main effect of response type, F(2,108) = 3.38, p = .04,pairwise comparisons showed that there was no differences between the responses (all z < 1.95, all p > .15).

Table 2
Descriptive statistics characterizing participants analyzed for Experiment 1, 2 and 3. Note: Erroneous values for age (age = 200) and completion time (time >120 min) were suppressed when calculating mean and standard deviation to ensure accurate estimation.

Experiment 2
Four participants committed a single response-outcome question error during the training phase.As such, knowledge of responseoutcome contingencies did not vary across response type, F(1, 66) = 1.08, p = .30.

Experiment 1
As seen in Fig. 2A, across both transfer tests, S congruent showed increased responses relative to S incongruent , though the size of the cueing effect was larger during the choice test than the alone test.This was confirmed by a significant interaction between stimulus type and test type, χ 2 = 10.97,p < .001,and a main effect of stimulus, χ 2 = 32.94,p < .001,and a main effect of test type, χ 2 = 35.11,p < .001.The baseline response rate significantly affected responding during stimulus exposure, χ 2 = 15.87,p < .001,though there was no evidence for baseline responding interacting with stimulus or test type (all χ 2 < 1).After adjusting for baseline response rate, planned contrasts confirmed that significant specific transfer effects were observed in both alone, t( 875 both choice and alone transfer tests, but suggest that, when there is a choice between two rewarded outcomes, the response competition results in more pronounced inhibition of incongruent responses.

Experiment 2
As can be seen in Fig. 2C, there was evidence for specific transfer with greater responding during S congruent compared to S incongruent , χ 2 = 48.37,p < .001.Baseline response rate also significantly impacted the response rate during stimulus exposure, χ 2 = 122.61,p < .001,and there was no evidence for an interaction between stimulus effects and baseline rates, χ 2 = .65,p = 42.After controlling for baseline response rate, S congruent provoked higher responding compared to S incongruent , t(401) = 6.96, p < .001,d = .82.

Experiment 3
As expected, the response rate during stimulus exposure was significantly influenced by baseline response rate, χ 2 = 23.69,p < .001,though there was no interaction between stimulus condition and baseline response rate, χ 2 = .11,p = .74.After controlling for baseline response rate there was a main effect of stimulus type, χ 2 = 99.22,p < .001.S congruent provoked significantly more responding compared S in- congruent , t(1083) = 9.96, p < .001,d = .78.

Experiment 1
As seen in Fig. 2B, responding during the S similar was greater than during the S-which was greater than during S distinct across both alone and choice transfer tests, suggesting that response excitation induced by reward cues was dependent on similarities between the cue and the reward outcome earned by the response.There was a main effect of stimulus type, χ 2 = 22.59, p < .001,and no evidence for a main effect of test type, χ 2 = 2.60, p = .11,nor evidence for an interaction between the two factors, χ 2 = 1.22,p = .55.However, there was also a main effect of baseline response rate, χ 2 = 13.52,p < .001,and baseline response rate significantly interacted with stimulus type, χ 2 = 7.077, p = .03,and stimulus type by test type, χ 2 = 6.81, p = .03.There was no significant interaction with test type and baseline response rate, χ 2 = .57,p = .45.
After controlling for baseline response rate, planned contrasts in each test type revealed that, in the choice test, S similar provoked a higher response rate than S distinct , t(1315) = 5.83, p < .001,d = .72. S similar also provoked a higher response rate compared to the S-, t(1315) = 2.91, p = .008,d = .36,S distinct resulted in lower responding compared to the control stimulus, t(1315) = − 2.87, p = .008,d = .36.In the alone test, similar stimuli provoked a higher response rate than distinct stimuli, t (1315) = 4.68, p < .001,d = .55. S similar also provoked a higher response rate compared to S-, t(1315) = 2.92, p = .008,d = .34,but S distinct resulted in similar responding compared to the S-, t(1315) = -1.7,p = .09,d = .21.Further contrasts confirmed that, like with the incongruent condition in the specific transfer test, it was S distinct that was significantly impacted by the choice test, with response rate significantly higher during the alone test compared to the choice test, t(1315) = 2.89, p = .004,d = .36. S similar and S-both had comparable response rates in each test (Similar: t(1315) = 1.64, p = .10,d = .19;Control: t(1315) = 1.61, p = .11,d = .20).

Experiment 2
As seen in Fig. 2D, like the results of Experiment 1, response excitation to the reward cues was dependent on the similarity between the cue and the outcome earned by the response.Baseline response rates significantly impacted responding during stimulus exposure, χ 2 = 17.43, p < .001,but no interaction between baseline responding and the stimulus effects, χ 2 = .12,p = .94.After controlling for baseline response rate, there was a main effect of stimulus type, χ 2 = 44.08,p < .001,such that S similar provoked comparable responding to the S-, t(601) = 1.07, p = .29,d = .13,while S distinct provoked lower responding than the S-, t (601) = − 5.04, p < .001,d = .66.A further contrast confirmed that response rate during S similar was significantly higher than responding during the S distinct , t(601) = 6.32, p < .001,d = .80.

Experiment 3
As expected, response rate during stimulus exposure was significantly influenced by baseline response rate, χ 2 = 86.79,p < .001,After controlling for baseline response rate, there was a main effect of stimulus type, χ 2 = 46.41,p < .001.There was no evidence for an interaction between baseline response rate and stimulus effects, χ 2 = 3.51, p = .17.

What is the overlap between specific and general transfer effects?
Experiment 1 observed a large correlation between specific and category-specific general transfer, r(35) = .71,p < .001.Experiment 2 observed a medium correlation between specific and category-specific general transfer, r(32) = .43,p = .01.Experiment 3 observed a large correlation between specific and category-specific general transfer, r (89) = .51,p < .001.

Discussion
The results of the present study suggest that food cues increase foodseeking for rewards which are like the cued food in flavor profile but not for dissimilar foods.The experiments presented here found, in a general Pavlovian-instrumental transfer test, cues that represented rewards that were different from the rewards available via instrumental actions only provoked increased response rates when the cue represented foods that were similar in their flavor profile to the available response, and instead decreased response rates for instrumental actions that were associated with foods with distinct flavor profiles.As shown in Experiment 1, this was true when there is a choice of available actions for foods both similar and distinct to the cued food, but also when there is no choice of actions.
This finding is consistent with the food cue-reactivity literature.For example, Andersen, Byrne, and Wang (2023) asked participants to observe images of M&Ms and imagine eating them.Participants who received 3 trials showed an increase in their desire to eat.Specifically, this increase in the desire to eat was driven by a specific desire for something sweet, whereas the desire for salty or fatty foods did not change after 3 trials.This suggests that food cues prime desire for foods of a particular category (sweet/salty).Similarly, Ferriday and Brunstrom (2011) found that exposure to the sight and smell of highly palatable savory food (pizza) cued craving for other savory foods (pasta, eggs, chips and beans, and chicken tikka masala) but not for palatable sweet foods (cake and chocolate buttons).Importantly, these changes in craving were steeper in individuals with a higher compared to a lower weight.Our experiments are in line with these prior cue-reactivity studies and affirm that food cues invigorate reward-seeking for foods of a particular category, even when the food cue is representative of food that is not available for consumption.This highlights that people who might be struggling with their food choices (e.g., people with obesity or binge eating disorder) may be vulnerable to the impact of junk food cues (e.g., advertisements), even when the foods are not immediately available, if other foods with a similar flavor profile are available.However, an important limitation is that our convenience sampling did not assess a clinical sample such as individuals who are obese or have eating disorder pathology.Future research should ascertain the specific vulnerabilities of clinical populations.
Importantly, these experiments provide evidence for a goal-directed account of general transfer, where food choices are motivated by selecting the optimal action based on the expected availability and current desirability of a reward (Eder & Dignath, 2019;Hommel & Wiers, 2017;Mahlberg, Weidemann, et al., 2019;Shenhav, Botvinick, & Cohen, 2013).General transfer effects are typically considered to be due to energized responding that is produced via the incentive value associated with the reward cue, and the effect is also considered to impact actions non-specifically as the identity of the outcome associated with the general cue is thought to play no role in these transfer effects (Cartoni et al., 2016).Therefore, the energizing hypothesis would lead to the prediction that the two general cues in our experiments (S similar and S distinct ) would increase responding for all available reward-seeking responses to a similar extent.In contrast to this view, all experiments reported here suggest that the general food cues functioned as an occasion-setter, activating a goal for a particular food type (e.g., sweet foods), resulting in an increased response rate for an available reward-seeking action that was most likely to satisfy that goal (e.g., the chocolate response).Overall, there were strong correlations between specific and general transfer effects, suggesting both types of cue-reactivity are driven by overlapping processes.These results, however, stand in contrast to conventional associative learning theories.Cue-reactivity from general cues in a PIT paradigm is typically thought to provoke non-specific energization of reward-seeking.That is, the reward cue provokes motivation to respond, and this activation is independent of the particular perceptual qualities of the cued reward (i.e. the identity of the outcome), suggesting the energization effect is generalized across all available reward-seeking responses (Cartoni et al., 2016).To our knowledge, the experiments we report here are the first to assess the assumption of non-specificity that stems from the energization general transfer hypothesis.Typically, general transfer effects captured from PIT paradigms implemented with human participants are not designed to test for specificity in general transfer effects, instead analyzing cueing effects averaged across different response types.Given several reports of null results when testing for general transfer effects (Belanger et al., 2022;Geurts et al., 2013;Meemken & Horstmann, 2019;Nadler et al., 2011), our data provides an impetus for researchers to consider the selectivity of general transfer effects in future experiments.
An outstanding question is the degree to which the general transfer effects are mediated by reward desirability.Both the energization and the goal-directed hypotheses would posit that outcome desirability would contribute to the cue-reactivity effects observed in a general transfer test (Cartoni et al., 2016).However, where these hypotheses diverge is in predicting how selective changes in value for a particular type of reward (e.g. a sweet reward) would impact cueing effects in a general transfer task: an energization hypothesis would suggest devaluation of a sweet reward would impact the magnitude of general transfer from a savory food cue (and vice versa), whereas a goal-directed hypothesis would predict that devaluation would impact general cueing effects selectively, for food cues of a similar category.Therefore, future work is required to evaluate the validity of these two competing hypotheses of cue-reactivity during general transfer tests.Relatedly, an interesting future research question is whether devaluing the reward value of the food stimuli would moderate general transfer effects in a manner consistent with the goal-directed or energization hypotheses.
Our study included images as food cues rather than experimentally trained Pavlovian cues.The strength of this approach is its ecological validity.Images of foods provoke cue-reactivity that is associated with eating and weight gain to a similar degree as real foods (Boswell & Kober, 2016), and therefore provides the most compelling test of cue-reactivity for individuals struggling with their food choices.Moreover, Watson et al. (2016) found that images of foods provoke broadly similar cue-reactive effects to explicitly trained Pavlovian food cues, with the major difference being that Pavlovian food cues produced smaller and less reliable effects, which is likely due to the limited associative history participants have with the Pavlovian cues.However, future research should confirm these findings using explicitly trained Pavlovian cues, to ascertain whether cues with relatively short associative histories impact food-seeking in a similar manner.Finally, our first two experiments had small sample sizes largely because of the exclusion of a substantial portion of participants who showed degraded contingency knowledge, which might raise concerns for the reliability of these findings.However, despite the relatively smaller sample size we consistently observe both specific transfer as well as category-specific general transfer effects across the first two experiments, and in the final experiment which has greater statistical power and relatively low rates of degraded contingency knowledge.We also demonstrate that the effects persist when including participants with degraded contingency knowledge (and thus increasing sample sizes dramatically in experiment 1 and 2), which further suggests the effects are robust.Finally, our study did not assess characteristics such as obesity and eating disorder pathology but was a convenience sample of mostly student participants.Consequently, it is unclear the extent to which the effects observed here apply to these clinical profiles, therefore future research should ascertain the specific vulnerabilities of clinical populations.
In summary, the experiments reported here show that reward cues boost responding for particular rewards, including for other food rewards which are of a similar category (e.g., sweet) but not foods that are of a distinct category (e.g., savory).This provides evidence in favor of a goal-directed hypothesis for cue-reactivity observed in general transfer tests, contrasting conventional associative accounts that consider general transfer to be a non-specific energization of reward-seeking.These results provide important insights into the mechanisms that underpin the types of food cue-reactivity that result in increased food consumption and weight gain (Boswell & Kober, 2016).

Fig. 1 .
Fig. 1.Experimental blocks in Experiment 1. Note.The panels show screen shots of the experiment task used in Experiment 1.The green buttons were clickable, whereas the grey buttons indicate that the button was not clickable.Panel (A) shows selection of a response during the instrumental training and the awarding of a reward.Panel (B) shows selection of a response during instrumental training where only reward response is available and the awarding of a reward.Panel (C) shows the transfer test.The left section of (C) shows the baseline test, while the right section of (C) shows an example of stimulus exposure.The alone test blocks (not shown here) were identical, except with the middle button and either the left or right button active.Panel (D) shows an example of a question in the R-O knowledge test, which assessed R-O knowledge for all rewards.Experiment 2 and 3 are not shown here, but include blocks identical to A, C, and D with two exceptions: (1) both Experiment 2 and 3 included only two buttons (a left and right button), and (2) Experiment 3 varied the color of the buttons (left blue, right green) to improve R-O knowledge.

Fig. 2 .
Fig. 2. Results of Specific and General transfer tests in Experiment 1,2, and 3. Note.The panels show average response rate (per second) during stimulus exposure conditions where the central black line in the box represents the group average, the box represents the interquartile range, the line show the range of values, violin plot shows the distribution of responses, and the grey lines show median response rates connected across conditions for each individual in Experiments 1, 2 and 3. Panels (A), (C), and (E) shows the response rates in the test of specific transfer in Experiments 1, 2, and 3 respectively, with Panel (A) showing the results of the alone test in purple and the results of the choice test in green.Panels (B), (D), and (F) shows the response rates in the test of general transfer test in Experiments 1, 2 and 3, respectively.

Table 1
A summary of the task structure for the Pavlovian instrumental transfer task used in each experiment.
similar S distinctNote: (1.) Instrumental training in Experiment 1 included three blocks of training, order randomized between participants, which varied in the responses available.Each block was experienced once, and required participants earn 10 reward points to proceed.Instrumental training in Experiment 2 included two choice training blocks which required participants to earn 10 reward points in total in each block to proceed.Experiment 3 training followed Experiment 2, except that 20 reward points in total were required in each block to proceed.R salty > O chip represents the response-outcome contingency learned for the chip-associated response, and R sweet > O choc represents the response-outcome contingency learned for the chocolate-associated response.R 0 > X represents the response-outcome experienced for the responses that is not associated with a reward (only in 'alone' training blocks in Experiment 1).(2.)During transfer test(s), stimuli were presented while responses were available to press for rewards.S congruent and S incongruent were designated for testing specific transfer effects: S congruent that represented the same outcome that is associated with an available response (e.g. a chip image and R salty ); S incongruent denotes stimuli that represented a different outcome that is associated with an available response (e.g. a chip image and R sweet ); General transfer was tested with S similar, S distinct, and S-. S similar denotes stimuli that represented an outcome of a similar flavor that is associated with an available response (e.g. an ice-cream image and R sweet ); S distinct denotes stimuli that represented an outcome of a distinct flavor that is associated with an available response (e.g. an ice-cream image and R salty ); S-denotes stimuli that does not represent a rewarding outcome.R salty /R sweet represents the availability of both rewarded responses during the 'choice' test; R salty /R 0 and R sweet /R 0 shows that only one rewarded response and the response associated with no reward was available for each 'alone' test block in Experiment 1.randomized.Experiment 3 doubled the number of response-outcome tests, with a test after each training block, to ensure stronger R-O learning.